Origin Myths (II): Strategy and Interdisciplinary Life
Last night I watched my friend Daveed Gartenstein-Ross discuss counterrorism, PhDs and political science with Robert Wishart and Tom Nichols. Though I have many notes prepared as part of a longer series at Zenpundit on my emerging research interests, I'd like to offer these reflections on interdisciplinary life.
I ended up leaving a IR PhD program for something nebulous called "Computational Social Science." This is a decision that poses some risk. Not fitting into a "box" anymore yet possessing a substantive knowledge base that is mostly in international relations, comparative politics, security studies, and war studies I make it harder for myself when it comes to the guild-like structure of academia.
It's true (as I discussed in the linked Zenpundit post) that US political science was never really a great fit for me. This was not because of the usual reasons dissenters offer. I don't think poli-sci has too much math or jargon. I think political science is actually right where it needs to be, and is on the cusp of doing exciting things with groundbreaking new methodological tools. That said, I still chose to enter an intersdisciplinary program and leave IR.
As I noted in the Zenpundit post, no matter what I end up career-wise doing with PhD-enabled methods once I eventually wrap up my new PhD program, methodology can only be applied to what you already know. I have always strongly been inspired by strategy and military history most of all. I speak almost intuitively in the Clausewitzian canon that I learned for seven years. Hence I ultimately decided--after a lot of flirtations with different topics-- to stick with the literature I know and care about for my research program. Not only easier, but perhaps less feeling of schizophrenia.
Why didn't I just do it in IR/poli-sci? Well, strategy and political science have not really been really that tightly coupled since Schelling's Arms and Influence and The Strategy of Conflict. There are prominent exceptions, like the quantitative political violence field and game theorists that study conflict. But it's mostly disconnected from the strategy and war studies topics being pursued at places like King's College London's Department of War Studies or University of Reading's Strategic Studies program.
Daveed (as well as Yemen/HoA champion Gregory Johnsen) mentioned that topics that sit between applied and theoretical like counterterrorism are difficult to study within the mainstream US political science context, and I feel the same about strategy. Those who study strategy topics from the standpoint of what I'm familiar with are often not in the mainstream of the discipline, and are often more affiliated with policy institutions. As A.E. Stahl noted, strategy is something of an outcast in international relations (one of its natural homes)
The closest mainstream political science/international relations comes to thinking about strategy is heavily mathematical game theory and heavily qualitative-historical work on social movements like Tilly's Dynamics of Contention. But neither formal modeling nor middle range theory is exactly all of what I would need to do my work. Often times I felt fairly frustrated with the expansiveness of the theory I have at my disposal (books like Colin Gray's The Strategy Bridge or J.C Wylie's Military Strategy: A General Theory of Power Control) and the corresponding narrowness of the methods available.
I am often frustrated by the lack of methodological rigor that goes into strategic thought, and Lynn Rees' recent post on that subject goes to that point. But methodological rigor also doesn't mean one-dimensionality. Lately I have been corresponding a lot with fellow CSS PhD student Russell Thomas, who is doing his own fascinating research program on strategy and innovation in business. One of my favorite recent papers that Thomas sent me is a piece on the "frame problem" in artificial intelligence and cognitive science. It's not an new piece, it's an old problem that sits at the core of some foundational issues in AI. Daniel Dennett, a philosopher and cognitive scientist, is the author.
A certain passage leapt out at me. Dennett argues that it is not self-evident that humans learn from experience. If a child is spanked for taking from a cookie jar, how does the child then connect the idea of cookie-reaching to the idea of spanking, and then by extension to the idea of pain? In theory, as Dennett points out, instead of reacting to this perception of potential pain by refraining, the child could instead do any number of things (recite poetry, blink, sing, etc). At the time, Dennett argued that theorists have difficulty with this problem in philosophy because they take appearances for reality:
When one operates at the purely phenomenological or semantic level, where does one get one's data, and how does theorizing proceed? The term `phenomenology' has traditionally been associated with an introspective method - an examination of what is presented or given to consciousness. A person's phenomenology just was by definition the contents of his or her consciousness. Although this has been the ideology all along, it has never been the practice. Locke, for instance, may have thought his `historical, plain method' was a method of unbiased self-observation, but in fact it was largely a matter of disguised aprioristic reasoning about what ideas and impressions had to be to do the jobs they `obviously' did.
The myth that each of us can observe our mental activities has prolonged the illusion that major progress could be made on the theory of thinking by simply reflecting carefully on our own cases. For some time now we have known better: we have conscious access to only the upper surface, as it were, of the multi-level system of information-processing that occurs in us. Nevertheless, the myth still claims its victims. .....
So the analogy of the stage magician is particularly apt. One is not likely to make much progress in figuring out how the tricks are done by simply sitting attentively in the audience and watching like a hawk. Too much is going on out of sight. Better to face the fact that one must either rummage around backstage or in the wings, hoping to disrupt the performance in telling ways; or, from one's armchair, think aprioristically about how the tricks must be done, given whatever is manifest about the constraints. The frame problem is then rather like the unsettling but familiar `discovery' that so far as armchair thought can determine, a certain trick we have just observed is flat impossible.......
Dennett argues that the only reason this problem has been brought to the forefront is the practical problem of artificial intelligence:
Such utterly banal facts escape our notice as we act and plan, and it is not surprising that philosophers, thinking phenomenologically but introspectively, should have overlooked them. But if one turns one's back on introspection, and just thinks `hetero-phenomenologically' about the purely informational demands of the task - what must be known by any entity that can perform this task - these banal bits of knowledge rise to our attention. We can easily satisfy ourselves that no agent that did not in some ways have the benefit of the information (that beer in the bottle is not in the glass, etc.) could perform such a simple task. It is one of the chief methodological beauties of AI that it makes one be a phenomenologist in this improved way. As a hetero-phenomenologist, one reasons about what the agent must 'know' or figure out unconsciously or consciously in order to perform in various ways.
The reason AI forces the banal information to the surface is that the tasks set by AI start at zero: the computer to be programmed to simulate the agent (or the brain of the robot, if we are actually going to operate in the real, non-simulated world), initially knows nothing at all `about the world'. The computer is the fabled tabula rasa on which every required item must somehow be impressed, either by the programmer at the outset or via subsequent 'learning' by the system.
Dennett observes that AI programmers often get around this problem by trying to simply install everything an computational agent needs to know about the world it must operate in. But if the world changes, then the problem becomes both theorizing and designing a mechanism that is partly generative or productive that stores represents and stores information and ignores irrelevant information. Dennett goes on to talk about various shortcuts AI designers have made to try to get around this issue.
There are two things that strongly appeal to me about this passage in terms of my own research program.
First, Dennett's methodological discussions about phenomenologically-based theories and different modes of phenomenological inquiry brings to mind some core weaknesses with strategic theory. Bernard Brodie, in his 1949 article "Strategy as a Science," talked about how so-called "timeless" strategic concepts turned out to be not-so-timeless in World War II and instead were revealed to be shibboleth. This hasn't really changed.
A lot of discussion about strategy fits Dennett's idea of introspection -- the idea that strategic mechanisms can be divined from a naive study of history. History is valuable but as Colin Gray has noted again and again context remains king. Hence strategic theory often sits at a level of generality and (above all else) fuzziness about key terms, processes, and mechanisms that does not help actual investigation of strategy. There is a reason why Clausewitz talks about "critical analysis" as a structured way to study strategic history, and tries to define the nature of conflict as a precursor to analysis. Nature abhors a vacuum and people inevitably (and unconsciously) fill it.
Lynn Rees gets to this when he notes that strategic thought is forever trapped between vulgar new fashions ("complex adaptive"-something war) and appeals to venerable old authorities. Both to some extent function as what Dennett dubbed "cognitive wheels," elegant explanatory devices that are stand-ins for something that would explain how “a cognitive creature … with many beliefs about the world” can update those beliefs when it performs an act so that they remain “roughly faithful to the world."
Strategy is in large part how the strategist as cognitive creature bridges violence and political currency, producing new social facts with purposeful action. So there is a distinct similarity to the research problem. And while a cognitive wheel may be plausible in an abstract way, it also something of a dead-end as an argumentative tactic in terms of making theory.
The notion of cognitive wheels is not purely a problem for cognitive scientists and philosophers, but also for students of strategy as well. The sort of "timeless" shibboleths that Brodie lambastes, and in turn Rees discusses more generally are basically implicit cognitive wheels designed to stand-in for mechanisms, representations, and phenomenological reasoning about strategy's own "frame problems."
At the same time, the implications of what Dennett argues also is a problem not just for qualitative-historical strategy, but also for the kind of statistically based political science approach as well. Game theory and other modeling approaches are valuable to political science because it forces internal consistency on a theory, and what Dennett is talking about is not just consistency but also representation. These aspects should not be neglected. One concern that I do have about the emerging political science interest in data science tools is what one Evolutionary Game Theory blogger dubbed "learning without understanding" (when referring to different modes of machine learning methodologies).
I very strongly agree with the idea that prediction should be an aspect of theory. Without predictive value (especially out of sample prediction), a theory becomes as useful as the Norse "theory" that lightning was the work of an angry Thor. But it is easier to predict than to necessarily understand, and unless understanding is a core part of theory then we might just end up as more data-inspired and sophisticated versions of the Norsemen (and realist IR theorists) that modeler and data scientist Phil Schrodt rightly mocks in his marvelous piece on the poverty of quantitative political science.
This is where Dennett's emphasis on the hidden blessings of AI's trouble with "banal facts" comes to the fore. By having to start from zero and build a computational agent from the ground up, we have to make a host of mathematical, informational, and representational choices that become explicit. It makes one a "phenomenologist" in an "improved way," a "hetero-phenomenologist" that must reason about what the agent must know or figure out in order to perform a task.
Clausewitz's On War is ultimately a work that is strongly phenomenologist in that it reasons about what war as a gestalt entity must "know" or "figure out" (in the loosest of terms) to create the range of behaviors and outcomes that we must contend with today. This is why On War is ultimately so difficult for modern-day interpreters to understand. It represents a theoretical approach that modern social science has forsaken in many ways. Hence I see a natural home for classical strategic inquiry with the heirs of Alan Turing and Herbert Simon, not the aforementioned Bernard Brodie. Though, of course, Brodie and others of a similar persuasion do represent a powerful component of theory and methodology that Simon himself responded to at great length in his works.
The flaw in the approach that American social scientists took to studying strategy was precisely the problem of phenomenology and representation -- and this is what separates the King's College and Reading schools of strategy from the heirs of Brodie, Schelling, and Kahn. But the evolving study of computational social science, computer science, cognitive science, and artificial intelligence combines phenomenology with the statistical and formal approaches that characterize mainstream American social science. That's why I went to a different program.
Strategy is not going to ever be a science in the way Brodie desired. But also can focus with renewed vigor on the process of reason that Dennett describes in his essay. From philosophy, computer science, and cognitive science it could derive the hetero-phenomenologist approach, and from computer science's tools for building and experimenting on such ideas it could attain an very interesting experimental dimension as well.
Obviously I can do this most easily in an interdisciplinary program, and one with a very heavily computational backbone. And also one where computer science, software engineering, information systems, optimization, and AI classes outside of the department can be taken in. I am comfortable with the implication that it may make it much more difficult for me in terms of the academic market that I most fit (IR). However,the methodological tools and questions inherent in the AI and CS intersection with my likely future research interests are of value to entities beyond the academy. And the substantive areas of inquiry I want to pursue were unlikely to give me much mainstream IR disciplinary appeal anyway. So I feel that the trade is ultimately beneficial to me.
As always I will continue to revise my interests and research program based on new information and new thoughts. But I feel a fairly strong connection with this germ of an idea, and it is also much more practically feasible than learning an entirely new base of social science knowledge (and having to generate completely new questions to go along with it) then the one I already know and care about. Learning things in philosophy, CS, cogsci, and AI is hard. But those disciplines all to some extent reinforce each other and have a large overlap (if visualized as a Venn diagram). And they all fit with the core approach I've been interested in since my undergrad days.
So, if I were to ascribe the appeal of a interdisciplinary PhD in a TL: DR way, I'd say "the possibility of putting Clausewitz, Turing, and Brodie/Schelling all under one roof. What's not to like -- unless you are a Jominian player-hater?"
